Scaling rules reveal how vision language models combine images and text

Scaling Laws for Vision-Language Fusion

Machine LearningComputer Vision and Pattern Recognition

Summary

Vision-language models combine pictures and words so computers can understand both together. The authors studied two ways to train these models: mixing images and text early or combining them later after learning separately. They discovered a mathematical rule that predicts how well the model learns based on training steps, model size, and how the two types of data are combined. This helps decide which training method works best depending on the size of the model and the data available.

What this means in practice

  • For ai research teams: Guide decisions on training strategies for vision-language models depending on model size and data availability.
  • For machine learning engineers: Optimize multimodal fusion approaches to balance image understanding and text comprehension in large-scale AI systems.

Authors

Seng Pei Liew, Teppei Suzuki, Seiya Tokui

Abstract

Vision-language models (VLMs) have emerged as central foundation models combining multimodal perception with text comprehension. While VLMs are traditionally built via late fusion by continued pre-training an off-the-shelf language model on multimodal (image-text) data, mixing modalities from the start of pre-training (early fusion) has recently gained traction. However, the fundamental trade-offs between these paradigms remain poorly understood, resulting in expensive and heuristic-driven training recipes. In this paper, we systematically analyze the learning dynamics of these multimodal pre-training processes, showing that validation losses during vision-language fusion can be modeled by separating the base learning trajectory, where different data types contribute at distinct rates, from a distribution shift driven by the competition between modalities. We derive a scaling law that unifiedly describes these dynamics, enabling accurate loss predictions across mixture ratios, training steps, and model sizes. Crucially, our law characterizes the trade-off between early and late fusion: under a fixed data budget, early fusion preserves text performance while late fusion favors vision performance, yielding a trade-off that increasingly tilts toward early fusion as model size scales. Extensive experiments confirm that our scaling law reliably captures multimodal learning dynamics and governs fusion decisions.